63, P > 0.5). We relied on neurons that had spatial selectivity for the location of the stimuli, whose discharge rate was therefore informative

about the location of the salient stimuli, and with at least three error trials check details in the level 3 difficulty condition (Fig. 1D). A total of 63 neurons from dlPFC and 62 neurons from LIP satisfied these criteria and were used in this analysis. The time of target discrimination was computed for each area by comparing the responses to the salient stimulus in receptive field with distractors in receptive fields, using correct trials from stimulus presentations of difficulty level 1 (Fig. 2A and C) and level 3 (Fig. 2B and D). Consistent with a previous study from our laboratory that reported an early involvement of the dlPFC in bottom-up attention (Katsuki & Constantinidis, 2012a), the times of target discrimination were similar in this sample of neurons too, and in fact slightly earlier in dlPFC than LIP, for both level 1 stimulus (126 ms after stimulus onset in dlPFC, 133 ms in LIP) and level 3 stimulus

(171 ms in dlPFC and 183 ms in LIP). Behavioral outcomes were categorized into two groups, corresponding to correct and error trials. Only trials with lever errors following the match or non-match periods were identified as error trials for this analysis; errors CHIR-99021 cost due to breaks in fixation at any point, or releases of the lever before the offset of the stimulus, were excluded from analysis. Average firing rates of correct trials (dlPFC, 1140 trials; LIP, 1208 trials) and error trials (dlPFC, 525 trials; LIP, 832 trials) were plotted separately for each area (Fig. 3A and B). On average, the firing rates of error trials were lower than those of the correct trials in both dlPFC and LIP. To quantify the relationship between behavioral choices and neuronal responses, we performed a ROC analysis to compute the probability of distinguishing between the distributions of error and correct trials, involving identical stimulus

conditions, a quantity also known as choice probability (Britten et al., 1996), based on signal detection theory (see ‘Materials and methods’). The area under the ROC curve using the firing rate of correct trials Prostatic acid phosphatase and error trials represents the choice probability for each neuron. The choice probability was computed in a time-resolved fashion, in 250-ms windows, sliding in 50-ms intervals (Fig. 3C). The average dlPFC choice probability was significantly different from 0.5 for the cue and delay period (t-test; Cue, t62 = 5.15, P < 10−5; Delay, t62 = 4.25, P < 10−4), while significantly higher LIP choice probability than 0.5 was observed in all three task epochs (t-test; Fixation, t61 = 3.91, P < 0.001; Cue, t61 = 5.31, P < 10−5; Delay, t61 = 7.05, P < 10−8). A significant difference was present between areas in terms of choice probability.